An agent skill + CLI for analysing RenderDoc GPU captures (.rdc) without the RenderDoc UI: where frame time goes, NVIDIA Nsight Perf hardware counters, pipeline state and constant values, shader source/disassembly, before/after regressions, and shader-edit experiments measured in replay.
It runs on top of rdc-cli: each capture gets an rdc daemon that keeps it loaded, the skill's extraction and experiments run inside that daemon, and the agent can use rdc's own tools (shader debugging, pixel history, mesh export) on the same open capture. The .rdc is the only input; what the CLI extracts is kept in a working cache (%LOCALAPPDATA%\rdgpu\cache, Linux ~/.cache/rdgpu) so follow-up queries are instant.
Primary target: Windows + Unity (D3D11/D3D12/Vulkan captures, URP/HDRP/built-in markers, Unity debug pragmas) on NVIDIA GPUs. Linux works too.
Give your coding agent (Codex, OpenCode, …) this repository URL and ask it to install the skill. The agent should:
- Clone the repository and copy
.agents/skills/renderdoc-gpuinto the skills directory it reads (project or user level, e.g.<project>/.agents/skills/or~/.agents/skills/). - Run
node <skills dir>/renderdoc-gpu/scripts/rdgpu.mjs install. Without Node.js 18+ run the bootstrap instead — it puts a portable Node into the user profile (no admin) and then runsinstall: Windowspowershell -NoProfile -ExecutionPolicy Bypass -File <skills dir>\renderdoc-gpu\scripts\bootstrap.ps1, Linux/macOSsh <skills dir>/renderdoc-gpu/scripts/bootstrap.sh.install— this builds RenderDoc's Python module locally throughrdc setup-renderdoc: it needs git and a C++ toolchain (Windows: Visual Studio Build Tools with "Desktop development with C++") and takes 10–40 minutes. It installs uv and rdc-cli on the way and reports[NEED USER]for anything a person must do (admin prompt for Build Tools, NVIDIA login for the Nsight Perf SDK download). - Run
node …/rdgpu.mjs doctor --capture <some .rdc>and confirmSTATUS: ready.
Requirements: Node.js 18+ (installed by the bootstrap if missing), git, a C++ toolchain, network access to github.com and PyPI during install. Optional: the Nsight Perf SDK for NVIDIA hardware counters (setup-nvperf copies it into RenderDoc's plugin folder once downloaded). Without the rdc build the CLI falls back to an installed RenderDoc (qrenderdoc --python) with fewer features.
node .agents/skills/renderdoc-gpu/scripts/rdgpu.mjs doctor --capture frame.rdc
node .agents/skills/renderdoc-gpu/scripts/rdgpu.mjs open frame.rdc
node .agents/skills/renderdoc-gpu/scripts/rdgpu.mjs tree frame.rdc DrawOpaqueObjects --metrics ps,verts
node .agents/skills/renderdoc-gpu/scripts/rdgpu.mjs compare before.rdc after.rdc
node .agents/skills/renderdoc-gpu/scripts/rdgpu.mjs experiment frame.rdc 575 --variant fewer_taps=edit/v1/main.hlsl
session <capture> prints the rdc session name of a capture, so rdc-cli commands can be run on the same open capture (rdc --session <name> debug pixel <eid> <x> <y>). Run the CLI without arguments for every command. SKILL.md describes the workflow agents follow; references/ holds setup/troubleshooting, counter meanings, the analysis method, Unity debug pragmas, experiment guidance and answer formats.
| path | what |
|---|---|
.agents/skills/renderdoc-gpu/ |
the skill (SKILL.md, CLI in scripts/, references) |
scripts/bootstrap.ps1, scripts/bootstrap.sh |
first run without Node.js 18+: portable Node in the user profile (no admin), then install |
scripts/rdjob.py |
runs inside RenderDoc's Python (3.6-compatible): extraction, counters, state, shaders, experiments |
scripts/lib/*.mjs |
install, rdc-cli backend, host discovery, job runner, cache model, queries, compare, find-source, Nsight Perf setup |
tests/rdgpu.test.mjs |
node --test tests/rdgpu.test.mjs (offline); set RDGPU_TEST_CAPTURE=<.rdc> to also run a live replay test |
dev/gl-scene/scene.py |
synthetic URP-like OpenGL frame used to produce test captures (runs on Mesa llvmpipe under Xvfb) |
dev/fixtures/MiniUnityProject/ |
tiny Unity-like project for find-source tests |